Urban Traffic Monitoring with the Help of Bus Riders
Pengfei Zhou, Shiqi Jiang and Mo Li
School of Computer Engineering, Nanyang Technological University
{pfzhou, sjiang004, limo}@ntu.edu.sg
Abstract—Real-time urban traffic conditions are critical to
wide populations in the city and serve the needs of many
transportation dependent applications. This paper presents our
experience of building a participatory urban traffic monitoring
system that exploits the power of bus riders’ mobile phones. The
system takes lightweight sensor hints and collects minimum set of
cellular data from the bus riders’ mobile phones. Based on such
a participatory sensing framework, the system turns buses into
dummy probes, monitors their travel statuses, and derives the
instant traffic map of the city. Unlike previous works that rely
on intrusive detection or full cooperation from “probe vehicles”,
our approach resorts to the crowd-participation of ordinary bus
riders, who are the information source providers and major
consumers of the final traffic output. The experiment results
demonstrate the feasibility of such an approach achieving finegrained traffic estimation with modest sensing and computation
overhead at the crowd.
I. I NTRODUCTION
Real time urban traffic information is critical to wide populations living in the city. Comprehensive knowledge of instant
urban traffic conditions contributes to commuter’s better travel
planning, improved urban transportation and commuting efficiency, reduced road congestion and waste emission, as well
as other time and cost savings. For the past decades, increasing
efforts have been put into exploring an accurate, efficient,
and inexpensive way to instantly monitor the urban traffic
conditions. Conventional methods rely on intrusive sensing,
where people deploy infrastructural devices like magnetic loop
detectors or traffic cameras at roadsides to actively detect traffic references. Installing such intrusive devices incurs substantial deployment and maintenance costs and can only provide
limited observation at sparse positions. Recent studies resort
to the GPS traces collected from “probe vehicles” like taxis or
private cars to estimate the road traffic conditions [22], [25].
Such a passive probing method avoids cumbrous infrastructure
deployment and enjoys flexible information extraction from
the running probes in the city. Nevertheless, most of them
largely rely on full cooperation from the probe vehicles and
bear substantial cost in obtaining their location references.
In this paper, we describe our experience in building a
participatory urban traffic monitoring system. Our system
takes the operating buses as probe vehicles to sample the
road traffic conditions. Instead of requesting any GPS traces
from the transit agencies, we crowdsource the sensing jobs to
public bus riders using their commodity off-the-shelf (COTS)
mobile phones. The bus riders intelligently collect real time
traffic sensing data on buses and anonymously upload the
data. A backend server identifies and reorganizes the uploaded information from different buses, based on which
the travel time and average vehicle speed at different road
segments are estimated and a complete traffic map can be
finally generated. We primarily rely on the cellular signatures
together with the use of several lightweight sensing hints like
audio and acceleration signals from the mobile phones rather
than the energy expensive GPS data to derive the location
references. Compared with existing approaches, the proposed
traffic monitoring system provides a grassroots solution that
solely relies on the collaborative efforts of the public. Built
on the COTS mobile phones, our system obviates the need
for special hardware extension and is energy friendly, which
reduces operational overhead, encourages wide participation
and expands the service coverage.
The full implementation of such a participatory monitoring system, however, entails substantial challenges which
require practical solutions to cope with. First, tagging location
with cellular references is non-trivial. The cellular signals
themselves contain very rough location dependence. Precisely
locating moving vehicles with only cellular signals may suffer
from high localization error [16], [24] and huge overhead
for complete war-driving [21]. In this work, we present
a novel method which intellectually explores the invariant
location and cellular attributes of bus stops so as to build
a location mapping between the physical space and cellular
space. Second, the crowdsourced sensing data are complicated
and essentially carry error and noises. In this work, we
carefully treat the sensing data on both mobile phones and
the backend server. We do data cleaning at individual mobile
phones and develop clustering and aggregation method on the
backend server to process the joint data from all participants.
We consolidate above techniques and implement a prototype
system on Android phones and a laboratory server. During a
2-month experiment with 8 bus routes in a ∼25km2 region
in Singapore, we are able to collect the data input from 122
participants and derive instant traffic map of that region. The
experimental results demonstrate high accuracy in mapping
the traffic observations. The system power consumption is also
carefully examined.
The rest of this paper is organized as follows. We introduce
the background and motivation of urban traffic monitoring in
§II. The methodology and detailed system design are presented
in §III. The implementation and detailed evaluation results are
reported in §IV. We introduce the related work in §V and
conclude this paper and discuss possible future work in §VI.
Cumulative fraction
1
0.8
0.6
Stationary
Mobile on buses
0.4
0.2
0
0
20
40
60
80
100
120
GPS error (meters)
140
160
180
Fig. 1. Measurement of GPS localization errors in downtown Singapore.
II. BACKGROUND AND MOTIVATION
Numerous approaches have been proposed to monitor the
traffic conditions. Conventional traffic monitoring mainly relies on intrusive sensing infrastructures (e.g., the widely used
magnetic loop detectors, roadside cameras and speed meters)
to measure the spot speed of vehicles. The intrusive approach,
however, suffers from two major drawbacks. First, due to the
high implementation and maintenance cost, the systems are
usually sparsely deployed and provide limited coverage. For
example, a single loop detector costs $900∼2000 depending
on its type [19]. Second, measuring spot vehicle speeds at
certain places may not accurately capture the travel delays
along the whole road segments and introduce noises caused
by traffic interruptions and congestions.
In order to overcome such drawbacks, many studies resort
to using GPS traces from “probe vehicles” and measure the
average travel time to derive the full traffic map. Modern
public transport services cover most parts of the urban area
[1], [2], [3], which provides readily available probe vehicles
(e.g. numerous taxis) [18], [25]. Such an approach, however,
is strongly dependent on the cooperation from the transit agencies and requires installation of real-time Automatic Vehicle
Location (AVL) systems, which usually comes with substantial
cost. With a fleet of thousands of vehicles, the installation of
in-vehicle AVL systems incurs tens of millions of dollars [20].
In this paper, we leverage the public buses to probe the
traffic conditions, but make use of bus riders’ mobile phones,
and fundamentally decomposes the individual “probing” tasks
from the running buses. By doing so, our system does not
rely on the cooperation of any transit agencies or even any
particular probe vehicles. The crowd-participation of the bus
riders themselves contribute the source sensing data and consume the final traffic output. We believe such an approach
can be thus easily adopted to a wide range of different cities
with slight modification. This work is fundamentally different
from our previous work [27] on bus arrival prediction that the
problem of traffic monitoring looks at holistic data collection
and organization from unsorted bus trips rather than classifying
and inspecting specific bus routes in [27]. In this paper, we
focus on data fusion across bus riders’ updates for traffic
mapping and system scalability to support wider monitoring
field. Instead of using GPS data, our system primarily relies
on the energy-friendly and widely available sensing hints from
the phones to intelligently track vehicles (e.g., cellular signals).
No special hardware implementation or infrastructure devices
are needed. The energy efficiency of sensing resources and
automatic traffic data collection of the system brings negligible
overhead to the bus riders and their phones, which as a
result encourages user participation. There are some works
[4], [20], [22], [25] that explore the GPS traces of COTS
mobile phones to track vehicles or human movement. In this
paper we do not employ GPS due to the following two major
disadvantages. First, GPS suffers from big localization error
in the downtown streets. To understand the magnitude of GPS
tracking error, we perform a measurement study in downtown
Singapore and summarize the GPS errors in Figure 1. We
experiment with HTC sensation mobile phone and measure
the GPS errors stationary or moving on buses by calculating
the distance between the GPS position and the ground-truth
position. The median errors are as high as 41m and 68m,
respectively, and their 90th percentiles errors are 75m and
130m, respectively. Such big error is due to the complicated
immediate surroundings in the downtown area, where the high
buildings block the line-of-sight paths to GPS satellites and
cause multipath problem. It is made worse when the phones
are inside buses and the GPS signal is further attenuated.
Similar results are also observed in other works [12], [20],
[26], [17], which indicates that such a phenomenon is common
for many cities across the world. Second, GPS device is known
energy aggressive. We measure the energy consumption of
the GPS receivers on Google Nexus One mobile phones with
the Monsoon power monitor. Continuous GPS tracking incurs
as high as 300mW energy consumption (details in §IV-D).
Due to the limited battery capacity of COTS mobile phones,
people usually turn off the GPS module to save power, which
discourages user participation.
III. S YSTEM DESIGN
In this section, we first present the design methodology after
which we detail the practical techniques in system design and
implementation.
A. Design methodology
Bus routes have high coverage of the urban road systems.
For example, the bus route coverage ratio is as high as 75% in
London [2] and 79% in Singapore [1]. Figure 2(a) depicts the
Jurong West area with a size of ∼25km2 in Singapore, where
∼80% roads are covered by more than 20 bus routes. If we
can track the moving trajectories of buses, we are then able
to map down the probed traffic conditions in the region.
Solely using the cellular signals, however, is difficult to
instantly track the buses. In urban area, the coverage of a
typical cell tower is about 200∼900m, which cannot provide
adequate location references. However, the inherent constraint
of bus operation provides us a unique angle, i.e., buses strictly
follow determined routes and stop at known bus stops. As
Figure 2(a) depicts, more than 100 bus stops densely distribute
in the region and separate the road systems into small road
segments. The precise locations of the bus stops and detailed
bus route operations are public information which is readily
1
1
Bus route 179
Bus route 243
Bus route 252
Bus route 199
Bus route 257
0.9
0.8
0.7
0.7
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
1000 m
4000 ft
(a) Measured bus routes.
0.8
CDF
CDF
0.6
Bus stop
0.9
Bus route 179
Bus route 199
Bus route 243
Bus route 257
Bus route 252
0
1
2
3
4
Similarity
5
6
7
(b) Similarity of the fingerprints collected at
the same bus stop.
0
0
Overall CDF
Effective CDF
1
2
3
Similarity
4
5
6
(c) Similarity of the fingerprints of different
bus stops.
Fig. 2. Similarity measurement of bus stop fingerprints.
11605,10291,10293,
3891,13484
11121,13486,3893,
11122,11123
Bus stop
715,712,714,711,
11423,11601,716
10716,3705,21705,
3704,3702,3703
3892,10291,
13484,3893
3891,3893,3892,
13484,10291
11601,13096,11605,13484,
10291,10293,11606
3892,13353,11122,
13021
13484,10291,3891,
11605,15081
11605,10291,10293,
3891,13484
3892,13021,10291,
3891,13353
11605,10291,10293,
3891,13484
3891,10293,15081,
13092,11605,3892
13096,10291,13091,13094,
15081,13092,13095
13096,13092,13091,
13094,13095
250 m
1000 ft
Fig. 3. An example area with the cellular fingerprints of 15 bus stops.
available on the web. Therefore, if we can track the stopping
statuses of buses along the bus stops, we can naturally map
their moving trajectories and derive the traffic estimations on
the road segments in between bus stops. Assembling the traffic
estimations of all segments gives us the whole traffic map. In
order to implement such an idea, we consider fingerprinting
the cellular signals at different bus stops and mapping them
into the cellular space. Later we will be able to match those
bus stops in cellular space according to the cellular fingerprints
collected from the bus riders’ mobile phones.
To understand the practical feasibility, we conduct a set
of experimental studies to know how effectively the cellular
signals can be used to distinguish different bus stops. We
measure the cell tower signals at 86 bus stops of 5 bus routes
(bus route 179, 199, 243, 252, and 257 in the region as
shown in Figure 2(a)). We collect the cellular signals in two
situations: when we stand at the bus stop and when we pass
by the stops on a bus. Taking the time and weather factors
into consideration, we collect the cell tower signals on days
of different weather conditions and at different time of a day.
The mobile phone normally can capture the signals from
multiple cell towers at one time, and chooses to connect to
the one with the strongest signal strength. Typically there are
4∼7 visible cell towers at each bus stop in our experiment.
We order their cell IDs according to their Received Signal
Strengths (RSS) and use such an ordered set to signature each
bus stop in cellular space. Figure 3 depicts an example area
where the cellular fingerprints of 15 bus stops are measured.
For each bus stop, we collect the set of all visible cell towers
and rank them in descending order of the RSS. The sets of
cell IDs for different bus stops are highly different from each
other. The bus stops well segment the road network.
We statistically analyze the similarities of the cell ID sets
collected at the same bus stop (sef-similarity) in different
runs. We use a matching algorithm (§III-C1) to calculate the
similarities, where higher scores represent higher similarities.
Figure 2(b) depicts the CDF of the self-similarity scores for all
bus stops of the 5 routes. We see that the overall self-similarity
score is very high. Generally 90% of the similarity scores are
higher than 3 and more than 50% of the similarity scores are
higher than 4, which demonstrates that the cell ID sets are
adequately stable to signature bus stops.
We also analyze the similarities of the cell ID sets collected
from different bus stops and plot the overall CDF of similarity
scores in Figure 2(c). We see that for over 70% of the bus
stops, their similarities are scored as 0 (no common cell IDs
at all) and for over 90% of the bus stops, they have similarity
scores lower than 2. We further examine the measurement data
and find that most of those similarity scores higher than 3 are
from the cell ID sets of two bus stops at opposite sides of
the two-way roads. In terms of location reference, they can
be treated as the same bus stop. Such treatment is proper and
does not degrade our system because the uploaded trip is timestamped, from which we can derive the bus moving direction
and map the traffic estimation to the correct side of the road.
We plot the effective CDF in Figure 2(c) with such treatment
and we see that more than 94% bus stops have similarity
scores lower than 2. The results suggest the feasibility of using
cellular signal fingerprints to distinguish different bus stops.
We choose cell tower signals over other possible wireless
signals to provide location inferences mainly due to the following considerations. First, mobile phones always maintain
connections to nearby cell towers to support telephone calls
and SMS service. The marginal energy consumption of collecting cell tower signals is negligible. Frequent scanning of other
Online/offline
Online data collection
Cellular
samples
Bus stop
database
Traffic
Monitoring
Update
Per sample
matching
Time stamps
Offline information
Bus routes
Upload periodically
Per bus stop
clustering
Traffic
model
Readily available
Per trip
mapping
Traffic
estimation
Fig. 4. System architecture and workflow.
wireless signals like WiFi, however, consumes much extra
energy [21]. Second, cellular network has almost complete
coverage of the entire city while other wireless signals usually
suffer from poor signal availability in many outdoor areas.
Third, the cellular signal sources are much more consistent
over time than other wireless signal sources like WiFi hotspots,
many of which are ad hoc and transient. The database built
on cellular signals is thus more stable and easier to maintain.
Based on such a methodology, we can effectively segment
the road network using bus stops and map down the traffic.
Figure 4 sketches the system workflow, concerning two major
components, i.e., online/offline data collection and trajectory
mapping for traffic estimation. The bus stop data database
can be updated in an online/offline manner. The real time
trips of bus riders are uploaded periodically. The offline bus
route and traffic model information is readily available. The
backend server carefully maps down the real time trips to
derive accurate traffic conditions. Each component will be
elaborated in the following subsections.
B. Data collection
As depicted in Figure 4 (top), the major system input
comprises three data sources.
Bus riders. The participatory bus riders serve as probes on
buses. The online collection of their trips starts automatically
when and only when users are detected on buses. We apply a
beep detection approach similar with that in our previous work
[27] to recognize the bus. Nowadays bus operators widely
employ IC card systems where most bus riders pay the fees by
tapping IC cards on the card readers which generate unique
beep response, e.g., EZ-link [1] in Singapore, Oyster [2] in
London, and MetroCard [3] in New York, etc. The beeps from
card readers are always made of audio signals of specific
frequencies (e.g., a combination of 1kHz and 3kHz audio
signals in Singapore and 2.4kHz in London).
We apply the Goertzel algorithm [5] to perform beep detection instead of FFT used in [27] to extract specific frequencies
(with prior knowledge of frequency components in the beep)
rather than all frequencies which significantly saves energy.
We measure and normallize the signal strength of several
typical frequency bands. If the signal strength of the 1kHz and
3kHz frequency band obviously jumps (an empirical threshold
of three standard deviation), we confirm the detection. We
use the standard sliding window averaging with window size
w = 300ms to filter out the noises and increase the robustness.
Once detecting the beep, the mobile phone starts recording
a trip. For each thereafter detected beep event, the mobile
phone attaches a timestamp and the set of visible cell tower
signals. The sensing data on the mobile phone thus record
a sequence of timestamped cellular samples in the trip. The
mobile phone concludes the current trip if no beep is detected
for 10 minutes, and starts uploading another independent trip
when new beeps are thereafter detected. We first primitively
filter out the noisy beep detections (e.g., the rapid train stations
use the same IC card readers) by thresholding the acceleration
variance (measured by the acclerometer) to distinguish the
people mobility pattern on rapid trains from taking buses.
It is based on the observation that buses usually move with
frequent acceleration, deceleration and turns, while rapid trains
are operated more smoothly.
Bus stop database. There is a database storing cellular
fingerprints of all bus stops which can be built online/offline.
The server relies on this database to identify the bus stops for
the uploaded cellular samples.
Bus routes and traffic model. The information of bus
operational routes is readily available from bus operators and
imposes constraints on how bus stops can be passed, which we
will exploit for trajectory mapping. There have been available
traffic models [10], [11], [18] giving the relationship between
the travel speeds of public buses and ordinary automobiles,
which we will adhere to for deriving the general traffic
conditions from buses.
For all bus stops, we aggregate the bus stops located at
the same location but different sides of the road as one,
and record their relative locations in the bus operational
routes. To yield the general traffic conditions from the bus
speed estimation, we make use of a traffic model [10] which
studies the relationship between public buses and ordinary
automobiles. We implement the model to our data set and
choose appropriate parameters according to our experiments
(§III-D).
C. Trajectory mapping
As depicted in Figure 4 (bottom), we do trajectory mapping
using bus stops as landmarks. We consider the received
sequence of cellular samples of each independent trip, based
on which the backend server identifies the passing by bus stops
and maps the trip trajectory down. Three levels of mapping
are done to refine the accuracy.
1) Per sample matching: With beep detection, the mobile
phone not only recognizes buses but also indicates the arrival
at bus stops because people only tap their IC cards at bus stops
for paying transit fees and the card readers are usually disabled
after buses move away from bus stops. Each uploaded cellular
sample thus corresponds to a particular bus stop.
We match each cellular sample from one trip to a signature
set stored in the fingerprint database and classify the sample
into one bus stop. Many algorithms, like k-means clustering,
have been used for fingerprint matching. In our system, the
cupload
cdatabase
Match
Score
1
2
3
4
1
7
3
1
×
3
+1 -0.3 +1 -0.3
5
5
5
+1
Match Gap Mismatch
3
1
1
2.4
Accuracy (%)
100
TABLE I
B US STOP MATCHING INSTANCE .
80
60
40
20
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
1.1 1.2 1.4 1.6 1.8
2
Chosen threshold
cellular samples at bus stops may be collected under different
conditions (e.g., on/off buses, different weather, etc.). While
the cell tower RSS values may vary, their rank always preserves. Thus we use the modified Smith-Waterman algorithm
[23] which focuses on the orders rather than the absolute RSS
value to score the similarity of different sets.
The backend server arranges the cell tower IDs in the set in
the descending order of each cell tower’s RSS. We denote
the cell tower set of a cellular sample e(x) as cupload =
{u1 , u2 , · · · , ul } ordered by the RSS of the l cell towers.
(i.e., si ≥ sj , 1 ≤ i ≤ j), where ui and si denote the cell
IDs of cell tower i and its RSS, respectively. We denote the
cellular fingerprint of a given bus stop b(y) in the database as
cdatabase = {d1 , d2 , · · · , dq } also ordered by the RSS, where
q is the set length. We match e(x) and b(y) by comparing
the similarity of the two sets. cupload typically has a different
length with cdatabase . For each cell tower set, Smith-Waterman
algorithm compares the segments of all possible lengths to find
out the optimal alignment with one cellular fingerprint in the
database, and uses a scoring system to weigh the value of
match, mismatch and gap.
We modify the Smith-Waterman algorithm settings to adapt
to our system. The performance of such a matching algorithm
is mainly determined by two factors, the set length and the
penalty cost for gaps and mismatches. In our system, the
set length of the fingerprint for each bus stop (the number
of cell towers) is about 4 ∼ 7 which is sufficient for the
matching algorithm. We vary the value of mismatch penalty
cost from −0.1 to −0.9 and simulate the matching accuracy.
Choosing −0.3 as the penalty cost gives the best result. Table
I shows an illustrative example where the uploaded cell tower
set is cupload = 1, 2, 3, 4, 5 and compared with cdatabase =
1, 7, 3, 5. The matching algorithm scores 2.4 by aggregating
3 matches, 1 gap and 1 mismatch. We denote Sim(e(x), b(y))
as the similarity score of one cellular sample e(x) and that of
an actual bus stop b(y). The server selects the bus stop with
the highest similarity score from bus stop candidates. To filter
out noisy reports, we only keep the cellular sample when its
highest matching score with the candidate bus stops exceeds
a threshold γ. We set γ = 2 according to the measurement
results in Figure 1(b). All cellular samples whose highest
similarity score is lower than 2 are discarded without further
processing. If there are more than one matched bus stop, the
one with a larger number of common cell IDs is selected. We
denote M (e(x), b(y)) = 1 if the matching result of e(x) is
b(y), and M (e(x), b(y)) = 0 otherwise.
2) Per bus stop clustering: When a bus arrives at a bus
stop, there are usually a number of passengers boarding
Fig. 5. Setting threshold ε.
and alighting giving multiple beeps, and multiple cellular
samples are taken. They will be received in a time sequence
at the server, allowing us information redundancy for better
reliability in identifying the correct bus stop. We co-cluster
the individual cellular samples according to their matched bus
stops and timestamps, and identify the bus stop for closely
clustered reports with more confidence.
For a sequence of m cellular samples E = {e1 , e2 , . . . , em }
with timestamps T = {t1 , t2 , . . . , tm }, their matching results
are a set of corresponding bus stops {b1 , b2 , . . . , bm } with
their similarity scores {s1 , s2 , . . . , sm }. In our co-clustering
algorithm, we denote the maximum possible similarity score
as s0 and the maximum possible time interval between two
cellular samples for the same bus stop as t0 . In our system
implementation, parameters s0 and t0 are set to 7 and 30
secs, respectively. For two samples ei and ej , we weigh their
matching relationship as
s0 −|sj −si |
, if bi = bj
s0
L(ei , ej ) =
0,
otherwise.
Adding the timestamp information, we put ei and ej into the
same cluster if
t0 − |tj − ti |
+ L(ei , ej ) > ε,
(1)
t0
i.e., two samples are classified into the same cluster if they are
detected close in time and have similar matching results. We
use a threshold parameter ε to verdict their relationship. We
vary ε from 0 to 2 with step length 0.1 and test the clustering
accuracy according to an experiment trial with bus route 243.
The result is plotted in Figure 5. If the threshold is too small,
cellular samples from different bus stops might be treated as
one cluster while if it is too big, the samples at the same bus
stop might be classified into different clusters. Nevertheless,
we find the result reasonably tolerates threshold selection. We
can get satisfactory clustering accuracy with ε = 0.3 ∼ 1. In
our later system implementation, we choose ε = 0.6.
Figure 6 shows an example with a sequence of cellular
samples collected from one trip. The backend server clusters
the cellular samples into 2 clusters corresponding to bus stop i
and j and extracts the arrival time and departing time at each
bus stop, which will later be used to estimate the bus traveling
speed (in §III-D).
After the cellular sample clustering, we get a sequence of
n clusters {C1 , C2 , . . . , Cn }. Each cluster Ci corresponds to
a bus stop. As the cellular samples in the same cluster may
t
Cellular fingerprint matching
t
Clustering
Duration at bus stop i
Travel time
estimation
Departing
point
Arrival
point
Cellular sample
Duration at bus stop j
Departing
point
Arrival
point
Bus stop i
Bus stop j
match different bus stops due to noises, at this moment each
cluster Ci is associated with a pool of potential candidate bus
stops as Figure 7 shows (though our experiments suggest that
for most clusters there is only one candidate in the pool).
3) Per trip mapping: The operation of bus routes largely
constrains the possible combinations and sequences the
bus stops can be visited. Considering such constraints,
we can further filter out the impossible bus stop candidates and map each cluster of cellular samples to a sole
bus stop. As shown in Figure 7, a sequence of n cellular sample clusters are outputted from the previous step.
In each cluster Ck (k = 1, 2, . . . , n), there are a total number of Ek samples {ek (1), ek (2), . . . , ek (Ek )}, and
Bk potential candidate bus stops {bk (1), bk (2), . . . , bk (Bk )}.
Each candidate
bus stop bk (i) is assigned a probability
Ek
j=1
M (ek (j),bk (i))
and an average
Ek
j=1 [M (ek (j),bk (i))·Sim(ek (j),bk (i))]
.
E k
j=1 M (ek (j),bk (i))
E k
similarity sk (i) =
Our goal is to find out a segment from one bus route or all
possible concatenation of multiple bus routes that best matches
the current trip and successively derive the most “correct” bus
stop for each sample cluster. For two actual bus stops x and y,
we denote their relationship as R(x, y) = 1 if y is “behind”
x in some bus route, which means that buses might arrive
at y after passing by x, and R(x, y) = 0 if x = y, and
R(x, y) = −1 for the rest. Since there are probably more
than 1 bus stop candidates for some clusters, we can get a
set of all possible
n bus stop sequences S = {S1 , S2 , . . . , SN },
where N = k=1 Bk . Each Sj is comprised of a sequence
of n bus stops {b1 (aj (1)), b2 (aj (2)), . . . , bn (aj (n))}. We use
maximum likelihood estimation to find the best matching
sequence
S ∗ = arg max {p1 (aj (1)) · s1 (aj (1)) +
Sj:1∼N
n
[pi (aj (i))
i=2
C2
b2(1)
b2(2)
C3
b3(1)
b3(2)
C4
b4(1)
b4(2)
b1(3)
...
b2(3)
...
b3(3)
...
b4(3)
...
...
Fig. 7. Bus stop identification with a cluster sequence.
Fig. 6. Bus stop clustering of an example trip.
pk (i) =
C1
b1(1)
b1(2)
(2)
·si (aj (i)) · R(bi−1 (aj (i − 1)), bi (aj (i)))]}.
In Equation (2), we weigh a sequence Sj using both the
probabilities pi (aj (i)) and average similarities si (aj (i)). The
best matching sequence S ∗ finally maps down the trajectory
of the trip, and determines the “most likely” bus stop for each
cellular sample cluster on the trajectory.
D. Traffic estimation
Based on the mapping result, the backend server estimates
the average travel speed on road segments divided by bus
stops. As illustrated in Figure 6, for each uploaded trip, we
are able to identify the passing by bus stops and extract the
arrival time and departing time of each bus stop. We denote
the arrival time at bus stop i as ta (i) and the departing time
as td (i). The travel time between i and j is thus estimated
as tij = ta (j) − td (i). In practice, the bus may not stop at
one particular bus stop if there is no bus rider boarding or
alighting, and thus the information at the bus stop is missing.
In such cases, our method automatically treats the combined
two adjacent segments as one.
The bus travel time may not directly yield the general traffic
conditions. Public buses have more frequent stops and usually
adhere to more strict speed limits. The relationship between
the transit buses and general traffic conditions has been studied
in transportation domain [18], [10], [11]. We use a linear traffic
model similar with that in [10] to estimate ATT from BTT:
ATT = a + bBTT,
(3)
road length
where a = f ree
travel speed , representing the average travel
time of an automobile when there is little or no traffic, and
b represents the effect of traffic congestion (as measured by
the running time of buses) on ATT. The value of b can
be determined using linear regression and our experimental
measurement suggests that the value of b lies within a narrow
range [0.13, 0.18] for most road segments. For simplicity, we
select b = 0.15 for all road segments.
When we consider the trip reports from massive mobile
phones, for each road segment, there are typically more than
one speed estimation. In our system, we use a Bayesian
method [18] to combine the initial estimation with new data
input. We denote the variance of the historic mean speed v̄0
as σ02 and the variance of new mean speed v̄ as σ 2 . Then the
updated speed estimation is normal with mean speed v̄new and
2
variance σnew
v̄new =
v̄0
σ02
1
σ02
+
+
v̄
σ2
1
σ2
,
2
σnew
=
1
σ02
1
+
1
σ2
.
(4)
The updating procedure follows Equation (4) and produces
sequential travel speed estimations. The updating procedure
Bus route 179
Bus route 243
Bus route 252
Bus route 199
Bus route 257
Bus route 241
Bus route 182
Bus route 30
1000 m
4000 ft
Fig. 8. 8 concerned bus routes in the ∼ 25km2 implementation area.
uses the inverse of the estimation variance to weigh the historic
estimation and the updated estimations. The travel speed
estimation on each road segment is updated with a period of
T . In our experimental implementation, we set T = 15 mins.
IV. I MPLEMENTATION AND EVALUATION
We implement a prototype system on the Android platform
with different types of mobile phones, and experiment with
real data from two resources, manual collection and participatory collection, over a 2-month period. In this section,
we introduce the experiment methodology and detailed results. we first present our experiment environment details and
methodology (§IV-A). We test and evaluate the performance of
bus stop identification (§IV-B) and show the traffic estimation
results during the experiment period (§IV-C) and compare the
estimation results with the official traffic data from the transit
angency. The system overhead is also carefully investigated
(§IV-D).
A. Experiment methodology
Mobile phones. We develop the data collection app with
Android phones in Android OS 4.0.3 with API version 15. We
do controlled experiments mainly with three types of mobile
phones, i.e., HTC Sensation XE, HTC Desire S, and Google
Nexus 4 and 5. All types of mobile phones can easily support
the light computation and sensing overhead in our app. As
our system is independent of platforms, we believe that the
proposed method can be easily implanted to other OS and
hardware platforms, such as Apple iOS and Windows Phone.
Backend server. We implement a backend server in Java
for our experiments. It is running on the DELL Precision
WorkStation T3500 with 6GB memory and Intel Xeon(R)
CPU W3565 @ 3.20GHz(4 CPUs). It provides database update
and receives data from the participants.
Experiment environment. The public bus transit system
serves millions of bus riders every day covering most parts
of Singapore [27]. It is commercially operated by two major
transit companies, SBS Transit [6] and SMRT Corporation
[15]. IC card systems are widely used for paying transit fees.
Multiple card readers are installed on each SBS or SMRT
bus for collecting fees. Figure 8 depicts the experiment region
TABLE II
B US STOP IDENTIFICATION ACCURACY.
Route total errors error rate 1 stop error 2 stops error
182 121
8
6.61%
5
2
30
58
3
3.45%
3
0
241
80
6
7.5%
5
1
199
93
5
5.38%
4
1
in Singapore. Public transit buses periodically run on more
than 20 bus routes covering most of the roads in this 7km ×
4km area. Our experiment primarily concerns 8 bus routes,
i.e, bus route 179, 199, 241, 243, 252, 257, 182 and partial
part of route 30. These 8 bus routes cover a major portion
of the road system in this area. We did extensive experiments
to study our system design feasibility and evaluate the system
performance. The performance of each system component was
carefully examined. The experiments started from Jan. 2013,
ended in Mar. 2014, and took more than 2 months in total.
Data collection. The data used in our system contain two
parts: the fingerprint database of bus stops and the real time
sensing data from the participatory mobile phones. In our
experiments, there are two data resources.
We manually collect the cellular fingerprints of the bus
stops on the 8 bus routes. For each bus stop, multiple cellular
samples are primitively collected and the sample with the
highest similarity with the rest samples is chosen as the
fingerprint and stored in the database. The cellular data used
in §III-A and the evaluation data for bus stop identification
used in §IV-B are also manually collected. The Land Transport
Authority (LTA) [7] of Singapore also provides us their traffic
data measured from the AVL reports of over 10,000 moving
taxis, which we take as ground truth in the experiment.
During the experiment, 122 participants, mainly undergraduate students and volunteers, contribute real time bus
information to the backend server. The data collection app is
installed in each participant’s mobile phone and uploads the
sensory data through WiFi or 3G network. In the first month,
we receive limited data from the participatory bus riders
due to their small number. The data concentrate on frequent
taken bus routes. In order to comprehensively evaluate the
system performance with wider participation, we encourage
(with vouchers) the participants to intensively take buses for
9 days in total and provide richer traffic data for our system.
The experiment results of both the sparse and intensive data
collection stages are shown in §IV-C.
B. Bus stop identification
The accuracy of bus stop identification for the 8 experimental bus routes is shown in Table II. In order to evaluate our
bus stop identification algorithm, we take buses to collect the
cellular signals at bus stops for 8 rounds. The cellular signals
in one of the 8 runs are used as the fingerprints stored in the
database. The rest 7 runs of data are used to identify the bus
stops. In Table II, we summarize the statistical results of the
bus stop identification error for 4 bus routes. The results of the
other 4 bus routes are similar. The bus identification error is
50
40
50
40
30
20
30
A
20
B
1000 m
4000 ft
1000 m
4000 ft
(a) 8:30AM.
(b) 15:00PM.
(c) Google Maps.
Fig. 9. 2 snapshots of the traffic map at different time of a day and the Google Maps traffic.
(a) Traffic estimation on segment A
3normal
40
2slow
20
1very
0
slow
0
80
1
2
3
4
5
6
7
8
9
In this section, we report our experiment results from the
participatory sensing data, and compare the results with the
official data provided by LTA.
Figure 9 depicts 2 snapshots of the traffic maps at different
time points (8:30AM and 15:00PM) on an experiment day
when we encouraged most participants to intensively take
buses. We show the travel speed of automobiles in 5 levels
as shown in Figure 9(a). The average moving speed is mainly
30-50 km/h. Traffic conditions in the studied area are spatially
different. For example, for the traffic condition at 8:30AM
shown in Figure 9(a), the highest moving speed is higher
than 50 km/h (left and bottom sides) while, in contrary, the
lowest speed is as low as 20 km/h (middle side). Meanwhile,
the distribution of the traffic at the 2 time points are also
different. The overall moving speed at 15:00PM (Figure 9(b))
is relatively higher. There are few road segments at 15:00PM
with travel speed lower than 20 km/h. The low-speed road
segments at 8:30AM are close to each other in 2 main roads in
the middle of Figure 9(a) where there are routine bus shuttles
between a university and a rapid train station every 15 minutes
every morning. As Figure 9 shows, although we only concern 8
bus routes, the coverage for the roads in the area is higher than
50% and they cover most major roads. The coverage ratio is
much higher than the Google Maps traffic for the area (Figure
9(c)). We believe that we can obtain more comprehensive
traffic conditions if more bus routes are concerned.
We compare our traffic estimation results with the official
traffic data that we acquire from LTA [7]. We pick 2 typical road segments (A and B, as shown in Figure (c)) for
comparison. Figure 10 compares the travel speed estimation
10
11
12
13
(b) Traffic estimation on segment B
14
15
16
17
40
0
0
18
4fast
3normal
60
Official traffic
Our estimation
Google Map indicator
20
C. Traffic estimation
4fast
60
0
Google Map indicator
80
Travel speed (km/h)
smaller than 8% for all the 4 bus routes. Bus route 241 has the
highest identification error and it has 13 effective bus stops.
In this experiment, we analyze 80 cellular sets collected from
its bus stops and 6 of them are mis-identified. In the misidentified cases, the results of 6 cases are 1 bus stop away
from the actual bus stop and only 1 case is 2 bus stops away.
The mis-identification cases have little impact on the system
performance. The high accuracy of bus stop identification
guarantees the accuracy of travel speed estimation.
2slow
1very
slow
0
9:30 2 10:30 4 11:30 6 12:30 8 13:30 10 14:30 12 15:30 14 16:30 16 17:30 18
Time
Fig. 10. Traffic estimation compared with official traffic and Google Map
indicator.
of automobiles (vA ) from our system, the travel speed (vT )
from official traffic data, and the traffic conditions indicated
from Google Maps on the two road segments for the time
period from 9:30AM to 17:30PM on another experiment day.
17 values are plotted for vA and vT respectively, each of
which is an average speed for a 15-minute window. The traffic
conditions from Google Maps are not the exact travel speeds
but 4 traffic levels indicated as very slow, slow, normal, and
fast. As depicted in Figure 10, Google Maps only provide
rough traffic levels which are not fine-grained in time and may
not accurately reflect the instant traffic conditions. vA and vT
are more sensitive to the traffic variation.
When we compare vA and vT on the two segments, they
are not always perfectly matching with each other but exhibit
interesting relationships. The speed estimates of vA are divided
into 3 groups, i.e., low speed (< 45 km/h), medium speed
(40∼50 km/h) and high speed (>50 km/h). When the travel
speed is low, vA perfectly matches vT . When the travel speed
is high, there is usually a gap between vA and vT . This is
probably because vA is the general traffic estimation derived
from the bus speeds, which are usually capped by lower
speed limits, while taxis on the other hand usually travel more
aggressively and yield a higher vT when the traffic is light.
Nevertheless, we clearly observe that vA follows the variation
pattern of vT for most of the time.
1
TABLE III
P OWER CONSUMPTION COMPARISON ( IN M W).
CDF
0.8
0.6
0.4
Medium−speed (40~50 km/h)
Low−speed (<40 km/h)
High−speed (>50 km/h)
0.2
0
0
2
4
6
8
10
12
Speed difference (km/h)
14
16
18
Fig. 11. Speed difference Δv between our traffic estimation vA and the
official traffic data vT .
In Figure 11, we statistically summarize the available
estimation results during the 2-month experiment, and plot
the CDF of the speed difference Δv between vT and vA
for all road segments and time durations when both are
available. We categorize the comparison cases into 3 types, i.e.,
high-speed traffics (vA > 50km/h), medium-speed traffics
(40 ≤ vA ≤ 50km/h) and low-speed traffics (vA < 40km/h)
and plot them separately. The majority of the studied cases
are of medium-speed traffics. As depicted in Figure 11, Δv
is the lowest (mostly about 3 ∼ 5) for low-speed traffics and
the highest (mostly about 8 ∼ 12) for high-speed traffics. For
medium-speed traffics, Δv is more disperse across 0 ∼ 12.
The results suggest that the estimated traffic speed from
our system generally provides a good measure of the traffic
conditions. It is particularly indicative for heavy traffics and
congestions that usually lead to low road speed.
D. System overhead
The computation complexity of the algorithms on mobile
phones is bounded by the Goertzel algorithm used for the
frequency extraction. The complexity of Goertzel algorithm is
O(Kg N M ) and that of FFT is O(Kf N log N ), where Kg and
Kf are the “cost of operation per unit” of the two algorithms,
respectively. M is the number of measured frequencies and
N is the sampling values. When the number of calculated
terms M is smaller than log N , the advantage of the Goertzel
algorithm is obvious. As FFT code is comparatively more
complex, the factor Kf is often much larger than Kg [5]. We
perform bus detection with microphone at the sampling rate
of 8kHz. By using the Goertzel algorithm instead of FFT, the
power consumption of the data collection app is reduced by
60mW.
We use Monsoon power monitor to measure the power consumption of two types of mobile phones (HTC Sensation and
Nexus One) under different sensor settings and summarize the
results in Table III. For each setting, we record the consumed
energy over a period of 10 mins. The mobile phone screen
is switched off during the measurement. The relative standard
deviation is also shown in the parentheses. We measure the
power consumption when no sensors are activated as a baseline
case. The power consumption of sampling cellular signals
is negligible for smartphones. We then measure the power
consumption of GPS tracking at a sampling rate of 0.05Hz,
Sensor settings
No sensors
Cellular 1Hz
GPS
Cellular+Mic(Goertzel)
GPS+Mic(Goertzel)
HTC Sensation
71(6)
72(6)
304(32)
182(20)
447(45)
Nexus One
84(5)
85(8)
333(41)
196(22)
443(57)
which is already considered very low for vehicle tracking [22].
The average power consumption as high as 304mW for HTC
and 333mW for Nexus One. The overall power consumption
of our data collection app is 182mW for HTC and 196mW
for Nexus One but it can be as high as ∼450mW if we use
GPS instead of cellular data to track the trips of bus riders.
V. R ELATED WORK
Traffic estimation. In transportation domain, some operational systems have been developed to measure the traffic
conditions using AVL system [10], [11], [18]. Chakroborty
et al. [10] study the possibility of using transit vehicles as
probes to predict automobile travel time. Pu et al. [18] propose
to use bus travel information to infer general vehicle traffic
conditions. Coifman et al. [11] use the transit fleet AVL data
to find all trips that use any portion of a prespecified freeway
segment. Researchers from computer science domain leverage
various location references to build up traffic monitoring
systems. VTrack [22] explores the GPS and WiFi references
using COTS mobile phones to tag the traffic estimations.
Aslam et al. [8] conduct a case study demonstrating that it
is possible to accurately infer traffic volume through GPS
data traces from a taxi fleet. Some works rely on heavy
deployment of static infrastructures providing very limited
spots of observations. For example, Kyun Queue [19] is a
sensor network system for real time traffic queue monitoring
with static sensor nodes deployed on roadsides. Our work
primarily differs from existing works. We rely on the bus
network to estimate the traffic conditions but fundamentally
decompose the “probing” tasks from the running buses. We
encourage participatory efforts from bus riders and derive the
traffic map without cooperation of any transit agencies.
Tracking and localization. Many approaches for vehicle
tracking and localization have been developed recently [14],
[20], [21], [22]. Jiang et al. [14] quantify the relative proximity
among physical objects and humans using “proximity zone”,
and compare the proximity zones created by various wireless
signals. Thiagarajan et al. [20] present a crowd-sourced alternative to conventional transit tracking system using GPS,
WiFi and accelerometer sensors on smartphones. CTrack [21]
presents energy-efficient trajectory mapping using celltower
fingerprints and utilize various sensors on mobile phones to
improve the mapping accuracy. This work primarily differs
from those works in its goal of urban traffic mapping. Our
approach does not necessarily rely on tracking or locating
individual bus routes. Accurate bus stop identification and
mapping suffices to project the traffic estimation down to the
map. The system scalability and deployment overhead is a key
design consideration under the crowdsourcing framework.
VI. C ONCLUSIONS AND FUTURE WORK
This paper presents the design and evaluation of a participatory urban traffic monitoring system. We leverage public
bus networks to estimate the traffic conditions, and explore
bus route constraints and bus stop references to generate the
traffic map. Our system relies on participatory efforts from bus
riders and utilizes lightweight sensing resources from COTS
mobile phones. We comprehensively evaluate the system with
a prototype system for 2 months in total in Singapore. The
implementation and experiment results demonstrate our system effectively monitors the traffic conditions with amiable
overhead. Due to its low cost and independence on any thirdparty cooperation, our system can be easily adopted to other
urban areas with slight modifications.
Future work includes deriving the overall traffic of a region
from the bus covered road segments. There have been some
existing models in transportation domain [9], [13], which can
be applied with our data feed. How to encourage bus riders’
participation for consistent and good performance is important.
At the initial stage, we may encourage the bus drivers to install
our app to bootstrap the system. Meanwhile we are putting up
the app on Google Play for possible experimental studies in
other areas and expect more user involvement.
R EFERENCES
[1] Bus transport in Singapore. http://en.wikipedia.org/wiki/Bus transport
in Singapore.
[2] Bus Transport for London. http://www.tfl.gov.uk/.
[3] New York city bus system. http://www.ny.com/transportation/buses/.
[4] Waze traffic and navigation. http://www.waze.com/.
[5] Goertzel algorithm. http://en.wikipedia.org/wiki/Goertzel algorithm.
[6] SBS Transit in Singapore. http://www.sbstransit.com.sg/.
[7] Land Transport Authority in Singapore. http://www.lta.gov.sg/content/
ltaweb/en.html.
[8] J. Aslam, S. Lim, X. Pan, and D. Rus. City-scale traffic estimation from
a roving sensor network. In ACM SenSys, pages 141–154, 2012.
[9] A. Bejan, R. Gibbens, D. Evans, A. Beresford, J. Bacon, and A. Friday.
Statistical modelling and analysis of sparse bus probe data in urban
areas. In IEEE ITSC, pages 1256 –1263, 2010.
[10] P. Chakroborty and S. Kikuchi. Using bus travel time data to estimate
travel times on urban corridors. Transportation Research Record,
1870:18–25, 2004.
[11] B. Coifman and S. Kim. Using transit vehicles to measure freeway
traffic conditions. In AATT, pages 90–101, 2006.
[12] W. Hedgecock, M. Maroti, J. Sallai, P. Volgyesi, and A. Ledeczi.
High-accuracy differential tracking of low-cost gps receivers. In ACM
MobiSys, pages 221–234, 2013.
[13] R. Herring, A. Hofleitner, and A. Abbeel, P.and Bayen. Estimating
arterial traffic conditions using sparse probe data. In IEEE ITSC, pages
929 –936, 2010.
[14] X. Jiang, C.-J. M. Liang, K. Chen, B. Zhang, J. Hsu, J. Liu, B. Cao, and
F. Zhao. Design and evaluation of a wireless magnetic-based proximity
detection platform for indoor applications. In ACM/IEEE IPSN, pages
221–232, 2012.
[15] F. Li, Y. Yu, H. Lin, and W. Min. Public bus arrival time prediction
based on traffic information management system. In IEEE SOLI, pages
336–341, 2011.
[16] K. Lin, A. Kansal, D. Lymberopoulos, and F. Zhao. Energy-accuracy
trade-off for continuous mobile device location. In ACM MobiSys, pages
285–298, 2010.
[17] D. Maier and A. Kleiner. Improved gps sensor model for mobile robots
in urban terrain. In IEEE Robotics and Automation (ICRA), pages 4385–
4390, 2010.
[18] W. Pu, J. Lin, and L. Long. Real-time estimation of urban street segment
travel time using buses as speed probes. Transportation Research
Record, 2129:81–89, 2009.
[19] R. Sen, A. Maurya, B. Raman, R. Mehta, R. Kalyanaraman, N. Vankadhara, S. Roy, and P. Sharma. Kyun queue: A sensor network system to
monitor road traffic queues. In ACM SenSys, pages 127–140, 2012.
[20] A. Thiagarajan, J. Biagioni, T. Gerlich, and J. Eriksson. Cooperative
transit tracking using smart-phones. In ACM SenSys, pages 85–98, 2010.
[21] A. Thiagarajan, L. Ravindranath, H. Balakrishnan, S. Madden, and
L. Girod. Accurate, low-energy trajectory mapping for mobile devices.
In USENIX NSDI, 2011.
[22] A. Thiagarajan, L. Ravindranath, K. LaCurts, S. Madden, H. Balakrishnan, S. Toledo, and J. Eriksson. Vtrack: accurate, energy-aware road
traffic delay estimation using mobile phones. In ACM SenSys, pages
85–98, 2009.
[23] M. S. Waterman and T. F. Smith. Identification of common molecular
subsequences. Journal of Molecular Biology, 147:195–197, 1981.
[24] J. Yang, A. Varshavsky, H. Liu, Y. Chen, and M. Gruteser. Accuracy
characterization of cell tower localization. In ACM Ubicomp, pages
223–226, 2010.
[25] J. Yoon, B. Noble, and M. Liu. Surface street traffic estimation. In ACM
MobiSys, pages 220–232, 2007.
[26] J. Zheng, Y. Wang, and N. L. Nihan. Tracking vehicles with gps: Is it
a feasible solution? In the 13th Annual Meeting of ITS, 2002.
[27] P. Zhou, Y. Zheng, and M. Li. How long to wait?: predicting bus arrival
time with mobile phone based participatory sensing. In ACM MobiSys,
pages 379–392, 2012.